anatomical regions of the heart are not as easy to di↵erentiate. They report a Dice similarity index for their segmentations of 0.89± 0.03 and 0.89 ± 0.01 respectively. The Dice similarity index is similar to Jaccard but it can be shown that
Dice(A, B) Jaccard(A, B)
Measured in Dice similarity the result of our algorithm (multFF+MRFext) is 0.9678±0.0100. The conclusion we can draw from this fact is that CTA (i.e. con- trast enhanced CT imaging) together with our method provides the best results for fully automatic pericardium segmentation known to the author. But since not all patients can undergo CTA-scans, future work would include evaluating this method on non-contrast CT images.
7.5
Review of Algorithm
Figure 7.1 shows an example of a successful segmentation (in fact a slice of a suc- cessful 3D segmentation). The MRF has the advantage that it directly corrects not allowed segmentations (as in Figure 7.2). In areas where the pericardium is visible it also has the ability to fit the boundary onto the true pericardium (Fig- ure 7.3). When the pericardium is not visible or if the multi-atlas estimation is not good enough this does not work and the segmentation will fail to represent the boundary accurately (Figure 7.4).
SliceID: 185. WL = −74. WW = 500.
Gold Standard Multiatlas Multiatlas+MRF
Figure 7.1: An example of a part of a good segmentation. The red boundary is the boundary of the gold standard. The green line represents the boundary of the segmentation estimated by the feature based multi-atlas segmentation. The blue line corresponds to the same feature based multi-atlas segmentation incorporated into the MRF.
CHAPTER 7. DISCUSSION
SliceID: 130. WL = −74. WW = 500.
Gold Standard Multiatlas Multiatlas+MRF
Figure 7.2: Here the MRF corrects the segmentation that is wrongly estimated to being in the lung cavity. Red is gold standard, green multi atlas segmentation and blue is the segmentation also using the MRF.
SliceID: 250. WL = −74. WW = 500.
Gold Standard Multiatlas Multiatlas+MRF
Figure 7.3: In this example where the pericardium is visible (the intensities are slightly brighter under the red line representing the boundary of the gold standard) we can clearly see the e↵ect of the intensity dependent boundary cost of the MRF.
7.5. REVIEW OF ALGORITHM
SliceID: 215. WL = −74. WW = 500.
Gold Standard Multiatlas Multiatlas+MRF
Figure 7.4: If the multi-atlas segmentation is not satisfactory and the peri- cardium is not clearly visible the segmentation will not represent the boundary accurately.
In this work we have been optimizing the algorithm against the Jaccard index. This is sometimes a problem since the pericardium is not in fact a closed object. We have the pulmonary veins and arteries, the aorta and inferior and superior vena cava that all are considerable objects that pass through the pericardium. When the expert made the Gold Standard labelings he more or less arbitrary cut through these objects. This is in general handled by our algorithm by the fact that the atlas registrations also provides a label propagated guess of how these areas should be closed. But we see one unexpected e↵ect where this can fail.
The images series are acquired in a consistent manner. But we see slight variations. Figure 7.5 shows a slice of a volume where the image series starts unusually high (i.e. more of the arteries and veins in the top of the image is visible). Since most of the images of the atlases are cut just above where the manual labeling starts there is little information in the atlas set of what the heart looks like above the labeling. We can see this e↵ect by the atlases being stretched far above the gold standard resulting in a miss aligned segmentation.
Our algorithm can actually be summarized well by Figure 7.5. It shows a slice of the image on which our algorithm performed the poorest results (Jaccard index 0.9002) and where most of the e↵ects discussed are visible.
CHAPTER 7. DISCUSSION
SliceID: 130. WL = −74. WW = 500.
Gold Standard Multiatlas Multiatlas+MRF
Figure 7.5: This figure shows a slice of the image on which our algorithm per- formed poorest. In a way it summarizes the advantages and drawbacks of our algorithm. The segmentation fails in the top of this slice due to the image stretching further above the pericardium relative to the other atlases. This has a large e↵ect on the Jaccard index but arguably not as profound on the fat measurements. To the right and to the bottom left the pulmonary veins and the inferior vena cava pass through the pericardium and we can see that our algorithm struggle to close these areas in a way consistent with the the experts delineation. On the bottom left the multi-atlas segmentation was not so accurate. The e↵ect of the MRF is clear where it successfully corrects the segmentation that passes through the lung and half of the fatty area to the left but not on the bottom left where the multi-atlas registrations were too far from the correct boundary.